Use of Windows For Harmonic Analysis With DFT

PROCEEDINGS OF THE IEEE, VOL. 6 6 , NO.
1 , JANUARY 1978
51
On the Use of Windows for Harmonic Analysis with the Discrete Fourier Transform
FREDRIC J. HARRIS,
MEMBER, IEEE
Ahmw-This Pw!r mak= available a concise review of data windaws pad the^ affect On the Of in the ' of 7 aoise9 m the ptesence of sdroag barmomc mterference. We dm call attention t a number of common o in be r p ~ c r h of windows den with the fdF ~ used transform. This paper includes a comprehensive catdog of data win-
-=
compromise consists of applying windows to the sampled data set, or equivalently, smoothing thespectral samples. The two operations to which we subject the data are sampling and windowing. These operations can be performed in either order. Sampling is well understood, windowing is less
related to sampled windows for DFT's.
I. INTRODUCTION
HERE IS MUCH signal processing devoted to detection and estimation. Detection is the task of determiningif a specific signal set is present in an observation, while estimation is the task of obtaining thevalues of the parameters describing the signal. Often the signal is complicated or is corrupted by interfering signals or noise. To facilitate the signal sets, theobservation is detection and estimation of decomposed by a basis set which spans the signal space [ 11. For many problems of engineering interest, the class of signals being sought are periodic whichleads quite naturally to a decomposition by a basis consisting of simple periodic functions, the sines and cosines. The classic Fourier transform is the mechanism by which we are able to perform this decomposition. By necessity, every observed signal we process must beof finite extent. The extent maybe adjustable and selectable, but it must be finite. Processing a finite-duration observation imposes interesting and interacting considerations on the harmonic analysis. These considerations include detectability of tones in the presence of nearby strong tones, resolvability of similarstrength nearbytones, resolvability of shifting tones, and biases in estimating the parameters of any of the aforementioned signals. For practicality, the data we process are N uniformly spaced samples of the observed signal. For convenience, N is highly composite, and we will assume N is even. The harmonic estimates we obtain through the discrete Fouriertransform samples of the associated (DFT) are N uniformly spaced periodic spectra. This approach is elegant and attractive when the processing schemeis cast as a spectral decomposition in an N-dimensional orthogonal vector space [ 21. Unfortunately, in many practical situations, t o obtain meaningful results this elegance be must compromised. One such
Manuscript received September 10, 1976; revised April 11, 1977 and September 1, 1977. This work was supported by Naval Undersea Center (now Naval Ocean Systems Center) Independent Exploratory Development Funds. The author is with the Naval Ocean Systems Center, San Diego, CA, and the Department of Electrical Engineering, School of Engineering, San Diego State University, San Diego, CA 92182.
11. HARMONIC ANALYSIS OF FINITE-EXTENT DATA AND THE DFT
Harmonic analysis of finite-extent data entails the projection of the observed signal on a basis set spanning the observation interval [ 1I , [ 3 I . Anticipating the next paragraph, we define T seconds as a convenient timeinterval and NT seconds as the observation interval. The sines and cosines with periods equal to an integer submultiple of NT seconds form an orthogonal basis set for continuous signals extending over NT seconds. These are defined as
sin [%kt]
O<t<NT.
We observe that by defining a basis set over an ordered index k, we aredefining the spectrum over a line(called the frequency axis) from which we draw the concepts of bandwidth and of frequencies close t o and far from a given frequency (which is related to resolution). For sampled signals, the basis set spanning the interval of NT seconds is identical with the sequences obtained by uniform samplesof the corresponding continuous spanning set up to the index N / 2 ,
sin [ 3 T ] =sin
[ 5 ]J
n = O , l , * . . ,N - 1
We note here that the trigonometric functions are unique in that uniformly spacedsamples(overan integer number of periods) form orthogonal sequences. Arbitrary orthogonal functions, similarly sampled, do not form orthogonal sequences. We also note that an interval of length NT seconds is not the same as the interval covered by N samples separated by intervals of T seconds. This is easily understood when we
U.S. Government work not protectedby U.S. copyright
52
PROCEEDINGS OF THE IEEE, VOL. 6 6 , NO. 1 , JANUARY 1978
Nth T - ~ e c Sample'
Fig. 1 . N samples of an even function taken over an NT second interval.
'
- 1 4 - 5 4 - 3 - 2 - 1
P*"odlC
e x t m w n Of
P w o d extmion of ~ m M Wp Y e
- 9 4 . 7 4 - 5 4 - 3 - 2 . 1
Fig. 2. Even sequence under DFT and periodic extension of sequence under DFT.
realize that the interval oveq which the samples are taken is closed on the left and is open on the right (i.e., [-)). Fig. 1 demonstrates this by sampling a function which is even about its midpoint and of duration NT seconds. Since the DFT essentially considers sequences to be periodic, we can consider the missing end point to be the beginning of the next period of the periodic extension of this sequence. In fact, under the periodic extension, the next sample (at 16 s in Fig. 1 .) is indistinguishable from the sample at zero seconds. This apparent lack of symmetrydue to the missing (but implied) end point is a sourceof confusion in sampled window design. This canbe traced to the early work related to convergence factors for the partial sums of the Fourier series. The partial sums (or the finite Fourier transform) always include an odd number of points and exhibit even symmetry about the origin. Hence much of the literature and many software libraries incorporate windowsdesigned with true evensymmetry rather than the implied symmetry with the missing end point ! We must remember for DFTprocessing of sampled data that even symmetry means that the projection upon the sampled sine sequences is identically zero; it does not mean a matching left and right data point about the midpoint. To distinguish this symmetry from conventional evenness we will refer to it as DFT-even (i.e., a conventionaleven sequence with therightend point removed). Another example of DFT-even symof a periodically metry is presented in Fig. 2 as samples extended triangle wave. If we evaluate a DFT-even sequence via a finite Fourier transform (by treating the + N / 2 point as a zero-value point), the resultant continuous periodic function exhibits a non zero imaginary component. The DFT of the same sequence is a set of samples of the finite Fourier transform, yet these samples exhibit an imaginary component equal to zero. Why the disparity? We must remember that the missing end point under the DFT symmetry contributes an imaginary sinusoidal 2 n / ( N / 2 ) to the finite transform component of period (corresponding to the odd component at sequence position N / 2 ) . The sampling positions of the DFT are at the multiples of 21r/N, which, of course, correspond to the zerosof the imaginary sinusoidal component. An example of this fortuitous sampling is shown in Fig. 3. Notice the sequencef(n),
Fig. 3. DFT sampling of finite Fourier transform of a DFT even sequence.
is decomposed into its even and odd parts, with.the odd part supplying the imaginary sine component in the finite transform.
111. SPECTRAL LEAKAGE The selection of a fite-time interval of NT seconds and of the orthogonaltrigonometric basis (continuous or sampled) over this interval leads to an interesting peculiarity of the spectral expansion. From continuum the of possible frelproject i quencies, only those which coincide with thebasis w onto a singlebasis vector; all other frequencies will exhibit non zero projections onthe entire basis set. This is often referred to as spectral leakage and is the result of processing finite-duration records. Although the amount ofleakage is influenced by the sampling period, leakage is not causedby the sampling. An intuitive approach to leakage is the understanding that signals with frequencies other than those of the basis set are not periodic in the observation window. The periodic extension of a signal not commensuratewithits natural period exhibits discontinuities at the boundaries of the observation. The discontinuities are responsible for spectral contributions (or leakage) over the entire basis set. The formsof this discontinuity are demonstrated in Fig. 4. Windows are weighting functions applied to data t o reduce the spectral leakageassociated with finite observation intervals. Fromoneviewpoint, the window is applied todata (as a multiplicative weighting) to reduce the order of the discontinuity at the boundary of the periodic extension. This is accomplishedby matching as many orders ofderivative (of the weighted data) aspossible at the boundary. The easiest way to achieve this matching is by setting the value of these derivatives to zero or near to zero. Thus windowed data are smoothly brought to zero at boundaries the so that the periodic extension of the data is continuous in many orders of derivative.
HARRIS: USE OF WINDOWSFORHARMONIC
ANALYSIS
53
Fig. 4. Periodic extension of sinusoid not periodic in observation interval.
From another viewpoint, the window is multiplicatively applied t o the basis set so that a signal of arbitrary frequency will exhibit a significant projection only on those basis vectors having a frequency close t o the signal frequency. Of course both viewpoints lead to identical results. We can gain insight into windowdesign by occasionallyswitchingbetween these viewpoints.
IV. WINDOWS AND FIGURES
OF
MERIT
Windows are used in harmonic analysis t o reduce the undesirable effects related to spectral leakage. Windows impact on many attributes of a harmonic processor; these include detectability, resolution, dynamic range, confidence, andeaseof implementation. We would like to identify the major parameters that willallow performance comparisons between different windows. We can best identify these parameters by examining the effects on harmonic analysis of a window. An essentially bandlimited signal f ( t )with Fourier transform F ( u ) can bedescribed bythe uniformly sampled dataset f ( n T ) . This data set defies the periodically extended spectrum F T ( u ) by its Fourier series expansion as identified as
We recognize (4a) as the f i t e Fourier transform, a summation addressed forthe convenience of its even symmetry. Equation (4b) is the f i t e Fourier transform with the rightendpointdeleted, and (4c) is theDFT sampling of (4b). Of course for actual processing, we desire (for counting purposes in algorithms) that the index start at zero. We accomplish this by shifting the starting point of the data N/2 positions, changing (4c) to(4d). Equation (4d) is the forward DFT. The N/2 shift will affect only the phase angles of the transform, so for the convenience of symmetry we will address the windows as being centered at the origin. We also identify this convenience as a major source of window misapplication. The shift of N/2 points and its resultant phase shift is often overlooked or is improperly handled inthe definition of the window when used with the DFT. This is particularly so when the windowing is performed as a spectral convolution. See the discussion on the Hanning window under the cos(' ( X ) windows. The question nowposed is, to what extent is the finite summation of (4b) a meaningful approximation of the infinite summation of (3b)? In fact, we address the question for a more general case of an arbitrary window applied to the time function (or series) as presented in
+F,(u) =
n=-m
w ( n T ) f ( n ~ ) (-junT) exp
(5 )
where
w ( n T ) = 0,
F(o)=
[-*IT
In1 >
5,
N even
f ( t )exp dt (-jut)
(34
and
F T ( ~ exp (+jut) d u / 2 = )
(3c)
=J-r/T
and where
IF(u)l = 0,
I u I 2 3 [27r/TI
Let us examine now the effects of the window on our spectral estimates. Equation (5) shows that the transform F,(u) is the transform of a product. As indicated inthe following equation, this is equivalent t o the convolution of the two corresponding transforms (see Appendix):
For (real-world) machine processing, thedata must be of f i t e extent, and thesummation of (3b) can only beperformed as a finite approximation as indicated as
+Nlz
or
F,(u) = F ( u ) W ( 0 ) .
f ( n r ) exp (-junT)
Fa(u)=
n =- N / 2
N even (4a)
Fb(u)=
(NlZ1-l
n=
f(nT)exp(-junT) ,
Neven (4b)
-N/2
Equation ( 6 ) is the key to the effects ofprocessing finiteextent data. The equation can be interpreted in two equivalent ways,which will be more easily visualized with the aid of an example. The example we choose is the sampled rectangle window; w ( n T ) = 1.0. We know W ( u ) is the Dirichlet kernel 141 presented as
N-1
Fd(wk) =
n =O
f(nn(-juknT), exp Neven
(4d)
54
PROCEEDINGS OF THEIEEE, VOL. 66, NO. 1, JANUARY 1978
k
Fig. 5. Dirichlet kernel for N point sequence.
Except for the linear phase shift term (which will change due to the N / 2 point shift for realizability), a single period of the transform has the form indicated in Fig. 5 . The observation concerning ( 6 ) is that the value of F,(w) at a particular w , say o = 00, is the sum of all of the spectral contributions at each w weighted by the window centered at wo and measured at w (see Fig. 6 ) .
A . Equivalent Noise Bandwidth
From Fig. 6 , we observe that the amplitude of the harmonic estimate at a given frequency is biased by the accumulated broad-band noise included in the bandwidth of the window. In this sense, the window behaves as a filter, gathering contributions for its estimate over its bandwidth. For the harmonic detection problem, we desire to minimize this accumulated noise signal, and we accomplish this with small-bandwidth windows. A convenient measure of this bandwidth is the equivalent noise bandwidth (ENBW)of the window. This is the width of a rectangle filter with the same peak power gain that would accumulate the same noise power (see Fig.7). The accumulated noise power of the window is defied as +nlT Noise Power = No
IW(412 dw/2n
-0
Fig. 6. Graphical interpretation of equation (6). Window visualized as a spectral Nter.
(8)
--AIL
F ( Q ) lsignal =
n
---+w
where N o is the noise power per unit bandwidth. ParseVal's theorem allows (8) to be computed by Noise Power = -
Fig. 7 . Equivalent noise bandwidth of window.
N O w2 (nT). T n
The peak power gain of the window occurs at w = 0, the zero frequency power gain, and is defied by Peak Signal Gain
W(0)=
n
w(nT)
( loa)
2
fiter is matched to one of the complex sinusoidal sequences of the basis set [3]. From this perspective, we can examine the PG (sometimes called the coherent gain) of the fiter, and we can examine the PL due to the window having reduced the data to zero values near the boundaries. Let the input sampled sequence be defined by (12 ) :
f ( n T ) = A exp ( + j o k n T ) + 4(nT)
Peak Power Gain = W 2 ( 0 ) =
[F
(12)
w(nT)]
(lob)
Thus the ENBW (normalized by NOIT, the noise power per bin) is given in the following equation and is tabulated for the windows of this report in Table I
where q ( n T ) is a white-noise sequence with variance 0:. Then the signal component of the windowed spectrum (the matched filter output) is presented in
w(nT) A exp (+joknT) exp (-jwknT) w(nT).
n
=A
(13)
B. ProcessingGain A concept closelyallied to ENBW is processinggain (PG) and processingloss (PL) of a windowed transform. We can think of the DFT as a bank of matched fiters, where each
We see that the noiseless measurement (the expected value of the noisy measurement) is proportional to the input amplitude A . The proportionality factor is the s u m of the window terms, which is infactthedc signalgain of the window. For a rectangle window this factor is N , the number of terms in the window. For any other window, the gain is reduced due t o the window smoothly going to zero near the boundaries. This
HARRIS: USE OF WINDOWS FOR HARMONIC ANALYSIS
55
TABLE I WINDOWS AND FIGURES MERIT OF

HIGHEST SIDELOBE LEVEL Id61 SIDELOBE FALL. OFF (dBIOCT) WORST COHERENT SCALLOP CASE 3.WB GAIN BW PROCESS LOSS (dB) (BINS1 LOSS (dB) OVERLAP CORRELATION (PCNTI
WINDOW
EQUIV. NOISE
6.WB BW
(BINS)
BW
(BINS)
75%0L
W%OL
RECTANGLE TRIANGLE
-13
-6
-12
1.oo
0.50
1.oo 1.33
0.89
1.28
3.92
1 .82
3.92 3.07
1.21 1.78
75.0 71.9
50.0
- 27
25.0
DE LA VALLEPOUSSIN TUKEY
- 53
-14
- 24
0.38
1.92
1.82
0.90
3.72
2.55
49.3
5.0
U=0.25 U=0.50 U = 0.75
-18
0.88
-15 0.63 -18
- 18 - 46 - 24
-6
1.10 1.22 -19 1.26 1.79 1.30

1.65 2;08
1.01 0.75 1.15 1.31 1.71 1.21 1.45 1.75

1.54
2.96 2.24 1.73 1.02 2.09 1.46 1.03

1
3.39 3.1 1 3.07 3.54

3.23 3.64 4.21
3.33 3.50
1.38 1.57 1.80 2.38 1 2.08 2.58
74.1 72.7 70.5 54.5 69.9 54.8 40.4 61.3 56.O 44.6 61.6 48.8 38.3 67.7 57.5 47.2 69.6 64.7 60.2 55.9 65.7 59.5 53.9 48.8
63.0 58.6 54.4
44.4 36.4 25.1 7.4 27.8 15.1 7.4 12.6 9.2 4.7 1.90 20.2 3.40 13.2 9.0
20.0 10.6 4.9
EOHMAN POISSON
0.41 0.44 0.32 0.25 0.43 0.38 -18 0.29 0.42 0.33 0.28 0.51 0.43 0.37
a = 2.0
a = 3.0
U=4.0
HANNINGPOISSON
-19 -24 -31

-35 -39
NONE
.w
-6
-6
-18 -18 -6 -6 -6
U = 0.5 U = 1.0
U=2.0
CAUCHY
16 1 1.73 2.02
.x
1.87
1.64
1.11 0.87 1.48 1.71 1.36
3.94
2.14 2.30 2.65
a=xo
a = 4.0 a = 5.0
-31 -35 -30
1.76 2.06 1.39

1.90
1.34 1.50
1.68
3.83 2.20 4.28 1.13 2.53

3.14 3.40 3.73 3.1 2 3.23 3.35 3.48 3.20 3.38 3.56 3.74 3.27 3.40 3.52 3.29 3.47 3.45
1.86
GAUSSIAN
a=2.5 U = 3.0 U = 3.5
-42 -55 -69

-50 -60 -70 -80 -46 -57 -69 -82
-6 -6
-6
1.33 1.55 1.79

1.33
1 1.64 1.25 0.94
.w
2.18 2.52 1.85 2.01 2.1 7 2.31 1 .w 2.20 2.39 2.57 2.07 2.23 2.36 2.1 3 2.35 1.81 2.72 2.19 2.44 2.44
DOLPHa=2.5 CHEEYSHEV a = 3.0 U = 3.5 U =4.0

KAISERBESSEL
0
0
0.53
0.48 0.45
0 0 -6 -6 -6 -6
1.39 1.51 1.62 1.73 15 0 1.65 1.80
1.44 1.55 0.42 1.65 1.43 1.57 1.71 1.83 1.49 1.59
1.70 1.44 1.25 1.10 1.46

1 .m
22.3 16.3 11.9 8.7 16.9 11.2 7.4 4.8 14.2 10.4 7.6 14.0 9.0 9.6 3.8
12.6
a = 2.0
U=2.5 U = 3.0
u=3.5
BARCILONTEMES
0.49 0.44 0.40 0.37 0.47 0.43 0.41 0.46 0.42 -18
1.02 1.93 0.89 1.34 1.18

1.05 1.69
a = 3.0 u = 3.5 U=4.0
-53 -58 -68 -51
-6 -6 -6
-6
1.56 1.67 1.77 1.57 1.73 1.71
EXACT BLACKMAN
1.52 1.68 1.66 1.90 1.56 1.74 1.74
1.33
62.7 56.7 57.2
BLACKMAN M I N I M U M 3-SAMPLE BLACKMAN-HARRIS ' M I N I M U M 4-SAMPLE BLACKMAN-HARRIS
- 58
-67 -92
-61
1.10 1.13 0.83 1.27 1.03

1.02
-6
0.42
-6
-6
0.36 0.45 0.40 0.40
2.00 1.61 1.79 1.BO
3.85
3.34
3.56
3.56
46.0
61 .O
'61 dE 3-SAMPLE
BLACKMAN-HARRIS
74 dB 4-SAMPLE
BLACKMAN-HARRIS 4-SAMPLE a = 3.0 KAISER-BESSEL
- 74
-69
-6
-6
53.9
53.9
7.4
7.4
*REFERENCE POINTS FOR DATA ON FIGURE 1 2 - NO FIGURES TO MATCH THESE WINDOWS.
56
PROCEEDINGS OF THE IEEE,
VOL.
j
I
66,
NO. 1 , JANUARY 1978
reduction in proportionality factor is important as it represents a known bias on spectral amplitudes. Coherent power gain, the squk of coherent gain, is occasionally the parameter listed in the literature. Coherent gain (the summation of (13)) normalized by its maximum valueN is listed in Table1. The incoherent component of the windowed transform is given by F ( w k ) lnois =
n
!:
r__-__
. \ ': ,
+Original <Windowed
' I
SequenTe" Sequences
., c----.,!
:
I ,
_---_. .. 1
i ? - O r i g i n a ls e q u e n c e
(Non-Overlapped) r n
:
. I
_ - - - -. . .'
i "
. 1
'
~,.--~-;~~-~
I
. :
, ,
Windowed Sequences (Overlapped)
w(nT)q(nT) exp ( - j w k n T )
(144
:,-
.,
and the incoherent power (the meansquare value of this component where E { } is the expectation operator) is given by
Fig. 8. Partition of sequences for nonoverlapped and for overlapped pro-g.

Region of Overlap = r N
E { I F ( W k ) Inois12} =
n m
w(nT) w ( m T ) E ( q ( n T ) q * ( m T ) }
(l-r)N
N-1
I
. exp ( - j w k n T ) exp (+jwkrnT)

= ui
n
w2(nT).
(14b)
N-1
rN-1
Fig. 9. Relationship between indices on overlapped intervals.
Notice the incoherent power gain is the sum of the squares of the window terms, and the coherent power gain is the square of the sum of the window terms. Finally, PG, which is defied as the ratio of output signalto-noise ratio to input signal-to-noise ratio, is given by
c
... ,
An important question related t o overlappedprocessing is what is the degree of correlation of the random components in successive transforms? This correlation, as afunctionof fractional overlap r , is defied for a relatively flat noise spectrum over the window bandwidth by (17). Fig. 9 identifies how the indices of (1 7)relate to the overlap of the intervals. The correlation coefficient
Notice PG is the reciprocal of the normalized ENBW. Thus large ENBW suggests a reduced processing gain. This is reasonable, since an increased noisebandwidthpermits additional noise to contribute to a spectral estimate.
C. OverlapCorrelation When the fast Fourier transform (FFT) is used to process long-time sequences a partition length N is first selected to establish the required spectral resolution of the analysis. Spectral resolution of the FFT is defined in (1 6) where A f is the resolution, f , is the sample frequency selected to satisfy the Nyquist criterion, and fl is the coefficient reflecting the bandwidth increase due tothe particular window selected. Note that [ f J N ] is the minimum resolution of the FFTwhich we denote as the FFT bin width. The coefficient fl is usually selected to be the ENBW in bins as listed in Table I
is computed and tabulated in Table I. for eachof the windows listed for 50-and 75-percent overlap. Often in a spectral analysis, the squared magnitude of successive transforms are averaged to reduce the variance of the measurements [ 5 ] . We know of course that when we average K identically distributed independentmeasurements, the variance of the average is related to the individual variance of the measurements by
-- _ _ OAvg. 2 1
2 K'
(18)
Af =
K If the window and the FFT are applied to nonoverlapping 50 percent overlap partitions of the sequence, as shown in Fig. 8, a significant 1 part of the series is ignored due to the window's exhibiting = - [ 1 + 2c2(0.75) + 2c2(0.5) + 2 c 2 ( 0 . 2 5 ) ] small values near the boundaries. For instance, if the transform K is being used to detect short-duration tone-like signals, the non 2 overlapped analysis could miss the event if it occurred near - 7 [c2(0.75) + 2c2(0.5) + 3c2(0.25)1, K the boundaries. To avoid this loss of data, the transforms are usually applied to the overlapped partition sequences as shown 75percent overlap. (19) in Fig. 8. The overlap is almost always 50 or 75 percent. This overlap processing of course increases the work load to cover The negative terms in (1 9) are the edge effects of the average and can be ignored if the number of terms K is larger than the total sequence length, but the rewards warrant the extra is ten. For goodwindows, ~ ~ ( 0 . 2 5 ) small compared to 1.0, effort.
fl(5).
Now we can ask what is the reduction in the variance when we average measurements which are correlated as they are for overlapped transforms? Welch [ 5 1 has supplied an answer to this question which we present here, for the special case of 50and 75-percent overlap
(16)
-= -
57
Nan RerolvaLde
PC&%
Resolvable Peaks
Fig. 10. Spectral leakage
effect of window.
Fig. 1 1 . Spectral resolution of nearby kernels.
and can also be omitted from (19) with negligible error. For this reason, c(0.25) was not listed in Table I. Note, that.for good windows (see last paragraph of Section IV-F), transforms taken with 50-percentoverlap are essentially independent.
D. Scalloping Loss An important consideration related to minimum detectable signal is called scalloping loss or picket-fence effect. We have considered the windowed DFT as a bankof matched fdters and have examined the processing gain and the reduction of this gain ascribable to thk window for tones matched to the basisvzctors. The basis vectors are toneswithfrequencies , equal to multiples, of f,/N (with f being the sample frequency). These frequencies are sample points the from spectrum, and are normally referred to as DFT output points or as DFT bins. We now address thequestion, what is the additional loss in processing gain for a tone of frequency midway between two bin frequencies (that is, at frequencies (k + 1/2)f,/N)? Returning to (1 3), with wlc replaced by w ( k +1 / 21, we determine the processinggain for this half-bin frequency shift as defined in
~ ( ~ ( 1 1 2 )/signal = )
be used. Additional comments poor on windows will be found in Section IV-G. We can conclude from the combined loss f i e s of Table I and from Fig. 12 that for the detection of single tones in broad-band noise, nearly any window (other than the rectangle) is as goodas any other. The difference between the various windows is less than 1.0 dB and for good windows is less than 0.7 dB. The detection of tones in the presence of other tones is, however, quite another problem. Here the window does have a marked affect, as will be demonstrated shortly.
F. Spectral Leakage Revisited
A
n
w(nT) ~ X (P i o ( l / z ) n ~ ) , where
qlI2)
1 ws 7r = - -=
2 N
NT'
(204
We also define the scalloping loss as the ratio of coherent gain for a tone located half a bin from a DFT sample point to the coherent gain for a tone located at a DFT sample point, as indicated in
Returning to ( 6 ) and to Fig. 6 , we observe the spectral measurement is affected notonlybythebroadband noise spectrum, but also by the narrow-band spectrum which falls within the bandwidth of the window. In fact, a given spectral component say at w = wo w contribute output (or will be l i observed) at another frequency, say at w = w, according to the gain of the window centered at 00 and measured at w,. This is the effect normally referred to as spectral leakage and is demonstrated in Fig. 10 with the transform of a finite duration tone of frequency wo This leakage causes a bias in the amplitude and the position of aharmonic estimate. Even forthecise of a single real harmonic line (not at a DFT sample point), the leakage from the kernel on the negative-frequency axis biases the kernel on the positive-frequency line. This bias is most severe and most bothersome for the detection of small signals in the presence of nearby large signals. To reduce the effects of this bias, the window should exhibit low-amplitude sidelobes far from the central main lobe, and the transition tothe low sidelobes should bevery rapid. One indicator of howwell a window suppressesleakage is the peaksidelobelevel(relative to the main lobe): another is the asymptotic rate of falloff of these sidelobes. These indicators are listed in Table I.
G . Minimum Resolution Bandwidth Fig. 11 suggests another criterion with which we should be concerned in the window selection process. Since the window imposes an effective bandwidth on the spectral line, we would be interested in the minimum separation between two equalstrength lines such that for arbitrary spectral locations their respective main lobes can be resolved. The classic criterion for this resolution is the width of the window at the half-power points (the 3.0-dB bandwidth). This criterion reflects the fact that two equalstrength main lobes separated in frequency by less than their 3.0-dB bandwidths will exhibit a single spectral peak and wiU not be resolved as two distinct lines. The problem with this criterion is that it does not work for the coherent addition we find in the DFT. The DFT output points are the coherent addition of the spectral components weighted through the window at a given frequency.
(20b) Scalloping loss represents the maximum reduction in PG due to signal frequency. This loss has been computed for the windows of this report and has been included in Table I.
E. Worst Case Processing Loss We now make an interesting observation. We define worst case PL as the sum of maximum scalloping loss of a window and of PL due to that window (both in decibel). This number is the reduction of output signal-to-noise ratio as a result of windowing of and worst case frequency location. This of course is related to the minimum detectable tone in broadband noise. It is interesting t o note that the worst case loss is always between 3.0 and4.3dB.Windows with worstcase PL exceeding 3.8 dB are very poor windows and should not
58
PROCEEDINGS OF THE IEEE,
VOL. 66, NO. 1 , JANUARY 1978
If two kernels are contributing to the coherent summation, the sum at the crossover point (nominally half-way between them) must be smaller than the individual peaks if the two peaks are to be resolved. Thus at the crossover points of the kernels, the gain from each kemel must be lessthan 0.5, or the crossover points must occur beyond the 6.0-dB points of the windows. Table I lists the 6.0-dB bandwidths of the various windows examined in this report. From the table, we see that the 6.0-dB bandwidth varies from 1.2 bins to 2.6 bins, where a bin is the fundamental frequency resolution wJN. The 3 .O-dB bandwidth does have utility as a performance indicator as shown in the next paragraph. Remember however, it is the 6.0-dB bandwidth which defies the resolution of the windowed DFT. From Table I, we see thatthe noise bandwidth always exceeds the 3.0-dB bandwidth. The difference between the two, referenced tothe 3.0-dB bandwidth,appears t o be a sensitive indicator of overallwindow performance. We have hs observed thatfor all the goodwindows on the table, t i indicator was found to be in the range of 4.0 to 5.5 percent. Those windows for which this ratio is outside that range either have a wide main lobe or a high sidelobe structure and, hence, are characterized byhighprocessingloss or by poor two-tone detection capabilities. Those windows for which this ratio is inside the 4.0 to 5.5-percent range are found in the lower left comer of theperformance comparison chart (Fig. 121, which is described next. While Table I does list thecommonperformance paramWORST W E PROCESSNG LOSS. dB eters of the windows examined in this report, the massof Fig. 12. Comparisonof windows: sidelobe levelsand worst case processnumbers is not enlightening. We do realize that the sidelobe ing loss. level (to reduce bias) and the worstcaseprocessingloss (to maximize detectability) are probably the most important parameters on the table. Fig. 12 shows the relative position defined as of the windows as a function of these parameters. Windows w ( n ) = 1.0, n = 0 , l ; - - , N1-. (21b) residing in the lower left comer of the figure are the goodwindow sequence is given in performing windows. They exhibit lowsidelobe levels and The spectral window for the DFT lowworst case processingloss. We urge the reader to read Sections VIand VII; Fig. 12 presents a lot of information, but not the full story.
V. CLASSIC WINDOWS
We will now catalog some well-known (and some not wellknown windows. For each window we will comment on the justification for its use and identify its significant parameters. A l the windows will be presented as even (about the origin) l sequences with an odd number of points. To convert the window to DFTeven, the right end point will be discarded and the sequence will be shifted so that the left end point coincides with the origin. We will use normalized coordinates with sample period T = 1.O, so that w is periodic in 2n and, hence, will be identified as 8 . A DFT bin wbeconsidered to l i extend between DFT sample points (multiples of 2nlN) and have a width of 2nlN.
A . Rectangle (Dirichlet) Window 161 The rectangle window is unity over the observation interval, and can be thought of as a gating sequence appliedt o the data so that they areof finite extent.The window fora finite Fourier transformis defined as
Thetransform of this window is seen to be the Dirichlet kernel, which exhibits a DFT main-lobe width (between zero crossings) of 2 bins and a first sidelobe level approximately 13 dB down from the main-lobe peak. The sidelobes fall off at 6.0 dB per octave, which is of course the expected rate for a functionwith a discontinuity. The parameters of the DFT window are listed in Table I. With the rectangle window now defmed, we can answer the question posed earlier: in what sense does the finite sum of (22a) approximatethe infinite s u m of (22b)?
+N 12
F(@ =
n=-N/2
f ( n ) exp (-in8)
(224
+F(8) =
n=--
f ( n ) exp ( - i d ) .
(22b)
w ( n ) = 1.0,
N n = --; - . , - l , O , l ; - - , ~ L
N
L
(2la)
is
and is shown in Fig. 13. Thesame-windowforaDFT
We observe the finite s u m is the rectangle-windowed version of the M i t e s m We recognize thatthe infinite s u m is the u . Fourier series expansion of some periodic function for which lo the f(n)'s are the Fourier series coefficients. We a s recognize
ANALYSIS
59
Fig. 13. (a) Rectangle window. (b) Log-magnitude of transform.
1.25
11,
4
(b)
OdB
-20
(a) Fig. 14. (a) Triangle window. (b) Log-magnitude of transform.
that the frnite sum is simply the partial sum of the series. From this viewpoint we can cast the question in terms of the convergence properties of the partial s u m s of Fourier series. From this workwe know the partial sum is the leastmeansquare error approximation t o the infinite s u m . We observe that mean square convergence is a convenient analytic concept, but it is not attractive for finite estimates or for numerical approximations. Mean-square estimates tend t o oscillate about their means, and do not exhibit uniform convergence. (The approximation ina neighborhood of a point of if more terms are added to the continuity may worse get partial sum.) We normally observe this behavior near points of discontinuity as the ringing we call Gibbs phenomenon. It is this oscillatory behavior we are trying to control by the use of other windows.
B. Triangle (Fejer, Bartlet) Window [7] The triangle window for a finite Fourier transform is defined as
and the spectral window corresponding t o the DFT sequence is given in
N InI W(n)=l.O-2' N/2' and is showninFig. defmed as

f
N n=-- .**,-l,O,l;--,14. The same window
(23a)
for a DFT is N
n = 0 ,l;..
W ( n )=
'(23b)
The transform of this windowseen is t o be the squared Dirichlet kernel. Its main-lobe width (between zero crossings) is twice that of the rectangle's and the firstsidelobelevelis approximately 26 dBdownfrom the main-lobe peak, again, twice that of the rectangle's. The sidelobes fall off at - 12 dB per octave, reflecting the discontinuity of the window residing in the first derivative (rather than in the function itself). The triangle is the simplest window which exhibits a nonnegative transform. This property canberealized by convolving any window (of half-extent) with itself. The resultant window's transform is the square of the original window's transform. A window sequence derived by self-convolving a parent window contains approximately twice the number of samples as the parent window, hence corresponds t o a trigonometric polynomial (its Z-transform) of approximately twice the order. (Convolving two rectangles each of N / 2 points will result in a triangle of N + 1 points when the zero end points are counted.) The transform of the window will now exhibit twice as many zeros as the parent transform (to account for the increased order of the associated trigonometric polynomial). But how has the transform applied these extra zeros available from the increased order polynomial? The self-
60
PROCEEDINGS OF THE IEEE, VOL. 66, NO. 1 , JANUARY 1978
of the cosine function. These properties are particularly attractive under the DFT. The window for a finite Fourier transform is defined as N
Fig. 1 5 . Two partial sums and their average.
w(n)=cosQ [ i n ] , and for aDFT as
N n = - - ***,-1,0,1;.* 2 2
w(n)=sinQ [ i n ] ,
n=0,1,2;*.,N-
1. (25b)
3 zf+1
Notice the effect due to the change of the origin. The most common values of a are the integers 1 through 4, with 2 being the most well known (sthe Hanning window). This window a is identified for values of a equal to 1 and 2 in (26a), (26b), (27a), and (27b), (the a for the finte transforms, the b for the DFT):
a = 1.O (cosine lobe)

N convolved window simply places repeated zeros at each location for which theparenttransform had a zero. This, of course, not only sets the transform to zero at those points, but also sets the firstderivative to zero at those points. If the intent of the increased order of polynomial is to hold down the sidelobe levels, then doubling up on thezeros is a wasteful tactic. The additional zerosmight better beplaced between the existing zeros (near the local peaksof the sidelobes) to hold down the sidelobes rather than at locations for which the transform is already equal to zero. In fact we will observe in subsequent windows that veryfewgoodwindows exhibit repeated roots. Backing up for a moment, it is interesting to examine the triangle window in terms of partial-sum convergence of Fourier series. Fejer observed that the partial s u m s of Fourier series were poor numerical approximations [ 8 ] . Fourier coefficients were easy to generate however, and he questioned if somesimple modification of coefficients mightlead to a new set with more desirable convergence properties. The oscillation of the partial sum, and the contraction of those oscillations as the order of the partial sum increased, suggested that an average of the partial sums .would be asmoother function. Fig. 15 presents an expansion of two partial sums near a discontinuity. Notice the average of the two expansions is smoother than .either. Continuing in this line of reasoning, an average expansion F N ( e ) might be defined by w(n)=cos [ i n ] , N n = -- **.,-l,O,l;*.,2 2 (26a)
a = 1.O (sine lobe)

w(n)=sin [in], n=0,1,2;*.,N1
(26b)
a = 2.0 (cosine squared, raised cosine, Hanning)
a = 2.0 (sine squared, raised cosine, Hanning)
=OS
1.0-cos
El1
-1
n = 0 , 1 , 2 , . . - , N - 1. (27b)
The windows are shown for a integer values of 1 through 4 in Figs. 16 through 19. Notice as a becomes larger, the windows become smoother and thetransform reflects this increased smoothness in decreased sidelobe level andfaster falloff of the where FM(6) is the M-term partial sum of the series. This is sidelobes, but with an increased width of the main lobe. easily visualized in Table 11, which lists the nonzero coeffiOf interest in this family, is the Harm window cients Of the first four sums and their (after the Austrian meteorologist, Julius Von Hann) [ 71. Not tion. We see that the Fejer convergence factors applied to the only is this window continuous, but so is its first derivative. Fourier series coefficients is, in fact, a triangle window. The sincethe discontinuity of this window resides in the second of Partial sums is known as the method Of cesko derivative, thebansform falls offat or at - 18 dB per summability. octave. Let us closely examine the transform of this window. C. CosQ(X)Windows We w gain some iteresting and l i insight learn of a clever application of the window under the DFT. This is actually a family of windows dependent upon the parameter a,with a normally being an integer. Attractions of The correct name of this window is a n The term Hanning Hn. the Wt which the terms Can be i used i this.report to reflect conventional usage. The derived term ih s n s generated, and the easily identified properties of the transform Hannd i also widely used.

0 dB
61
T
1.25
-.
(a) Fig. 16. (a) COS (nn/n? window.
(b) Log-magnitude of transform.

0 dB
't T
i
1.25
1.w
(a)
(b )
Fig. 17. (a) Cos2 (nn/N) win1dow. (b) Log-magnitude of transform.
1.25
(a) Fig. 18. (a) COS' (nn/N)window.
(b )
(b) Log-magnitude of transform.
I I
1.25 1.w
(a) Fig. 19. (a) Cos'
(b)
W / N ) window. (b) Log-magnitude of transform.
62
The sampled Hanning window can be written as the s u m of the sequencesindicated in
w(n) = 0.5 + 0.5
COS
[- I ,
n ,
1.(28a)
n = - -N * * - , - l , O , 1 , * . . , - 2 2
Each sequence hasthe easily recognized DFT indicated in

Fii. 20. Transform of Hanning window as a sum of three Dirichlet
kernels.
where
We recognize the Dirichlet kernel at the origin as the transform of the constant 0.5 samples and the pair of translated kernels as the transform of the single cycle of cosine samples. Note that the translated kernels are located on the f i t zeros of the center kernel, and are half the size of the center kernel. Also the sidelobes of the translated kernel are about half the size and of are opposite phase of the sidelobes of the central kernel. The summation of the three kernels sidelobes being in phase opposition, tends to cancel the sidelobe structure. This cancelling summation is demonstrated in Fig. 20 which depicts the summation of the Dirichlet kernels (without the phaseshift terms). The partial cancellingof the sidelobe structure suggests a constructive technique to define new windows. The most well-known of these are the Hamming and the Blackman windows which are presented inthe next two sections. For the specialcaseof the DFT, the Hanningwindow is sampled at multiples of 2n/N, which of course are the locations of the zeros of the central Dirichlet kernel. Thus only three nonzero samples are taken in the sampling process. The positions of these samples are at -2n/N, 0, and +2n/N. The value of the samples obtained from (28b)(including the phase factor exp (-j(N/2)0) to account for the N/2 shift) are - $, + - $, respectively. Notethe minus signs. These results , : from the shift in the origin for the window. Without the shift, the phase term is missing and the coefficients are all positive $, $. Theseare incorrect for DFT processing, butthey N find their way into much of the literature and practice. Ratherthanapplythe window as a product in the time domain, we always have the option to apply it as a convolution in the frequency domain. The attraction of the Hanning window for t i application is twofold; f i t , the window hs spectra is nonzero at only three data points, and second, the sample values are binary fractions, which can be implemented asright shifts. Thusthe Hanning-windowed spectral points obtained the from rectangle-windowed spectral points are obtained as indicated in the following equation as two real adds and two binary shifts (to multiply by
or as 2N real adds and 2N binary shifts on the spectral data. One other mildly important consideration, if the window is to be applied to the time data, is that the samples of the window must be stored somewhere, which normally means additional memory or hardware. It so happens that the samples of the cosine forthe Hanning window are already stored inthe machine as the trig-table for the FFT; thus the Hanning window requires no additional storage. D. Hamming Window /7] The Hamming.windowcan be thought of as amodified Hanning window. (Note the potential source of confusion in the similarities of the two names.) Referring back to Figs. 17 and 20, we note the inexact cancellation of the sidelobes from the summation of the three kernels. We can construct a window by adjusting the relative size ofthe kernels as indicated in the following to achieve a more desirable form of cancellation: w(n)=a+(l -a)cos
[$4
(304
Perfect cancellation of the f i t sidelobe (at 0 = 2.5 [2n/N]) occurs when a = 25/46 (aG 0.543 478 261). If a is selected as 0.54(anapproximation to 25/46), the new zerooccurs at 6 G 2.6 [ 2n/Nl and a marked improvement in sidelobe level is realized. For t i value of a,the window is called the Hamhs ming window andis identified by
3,
I
w(n) =
0.54
+ 0.46 COS
$n]
N n = - - *,-1,0,1,*.. 2 2
0.54 - 0.46 cos [ $ n ]
,
1. ( 3 0 ~
n=0,1,2;*.,N-
3: )
F(k)l H d n g =
3 [ F ( k )- 3 [ F ( k - 1)
+ F ( k + 111 1 I
R ~ (29) .
Thus a Harming window applied to a real transform of length
N can be performed as N real multiplies on the time sequence
The coefficients of the Hamming window are nearly the set which achieve minimum sidelobe levels. If a is selected to be 0.53856 the sidelobe level is -43 dB and the resultant window is a special case of the Blackman-Harris windows presented in Section V-E. The Hamming window is shown in Fig. 21. Notice the deep attenuation at the missing sidelobe position. Note a s that the small discontinuity at the boundary of the lo window has resulted ina l / w (6.0 dBper octave)rate of falloff. The better sidelobe cancellation does result in a much lower initial sidelobe level of -42 dB. Table I lists the param-
63
T T
1.2s
(a)
(b)
(a) Hamming window. (b) Log-magnitude of Fourier transform.
Fg 21. i.
1.25
(a)
(b)
Fig. 22. (a) Blackman window. (b) Log-magnitude of transform.
eters of this window. Also note the loss of binary weighting; hence the need t o perform multiplication to apply the weighting factors of the spectral convolution.
E. Blackman Window 171 The Hamming and Hanning windows are examples of windows constructed as the summation of shiftedDirichletkernels. This data window is defined for the finite Fourier transform in (31a) and for the DFT in (3 1b); equation (3 IC) the is resultant spectral window for the DFT given as a summation of the Dirichlet kernels D ( 8 ) defined by W ( 8 ) in (21c);
of this form with uo and u 1 being nonzero. We see that their spectral windows are summations of three-shifted kernels. We can construct windows with any K nonzero coefficients andachieve a (2K- 1) summation of kernels. We recognize, however, that one way to achieve windows with a narrow main lobe is to restrict K to a smallinteger.Blackmanexamined this window for K = 3 and found the valuesof the nonzero coefficients which place zeros at 8 = 3.5 (2n/N) and at 8 = 4.5 (2n/N),the position of the third and the fourth sidelobes, respectively, of the central Dirichlet kernel. exact These values and their two place approximations are
QO
=--
7938 10.426 590 71 N- 0.42 18608

L 0.496 560 62 N 0.50
(3 la)
9240 Ql=--
18608 18608
Q z =--1430 I0.076 848 67 N 0.08.
(31b)
Nl 2
m=O
The window which uses these two place approximations is known as the Blackman window. When describe we this window with the "exact"coefficients wewill refer to it as the exact Blackmanwindow. The Blackmanwindowisdefined for the finite transform in the following equation and (31c) the window is shown in Fig. 22:
W(n)= 0.42
Subject to constraint
Nf 2
Q ,
+ 0.50 cos
N n=--;..,-l,O, 2
1;** N - (32) ' 2'
= 1.0.
m=o
We can see that the Hanning and the Hamming windows are
64
PROCEEDINGS OF THE IEEE,VOL. 66, NO. 1 , JANUARY
1978
T T
1.25
If\
0)
(a)
Fig. 23. (a) Exact Blackman window. (b) Log-magnitude o f transform.
(a)
(b)
Fig. 24. (a) Minimum 3-term Blackman-Harris window. @) Log-magnitude of transform.
(a)
0)
Blackman-Harris window. (b) Log-magnitude of transform.
Fig. 25. (a) 44-
The exact Blackman window is shown in Fig. 23. The sidelobe sidelobe level. We have also constructed families of 3- and 4 level is 51 dB down for the exact Blackman window and is 58 term windows in which we trade main-lobe width for sidelobe dB down for the Blackman window. As an observation, note level. We call this family the Blackman-Harriswindow. We thatthe coefficients of the Blackmanwindow sum to zero have found that the minimum3-termwindowcanachievea (0.42 -0.50 M.08) at the boundaries while the exact coef- sidelobe level of -67 dB and that the minimum 4-term winof -92 dB. These windows ficients donot. Thus the Blackman window is continuous dow canachieveasidelobelevel with a continuous first derivative at the boundary and falls off are defiied for theDFT by like l/w3 or18 dB peroctave. The exact terms (Like the Hamming window) have a discontinuity at the boundary and w ( n ) = a o - a1 cos - n +a2 cos -2n - a 3 cos -3n , falls off like l / w or 6 dB per octave. Table I lists the parameters of these two windows. Note that for this class of winn = 0 , 1 , 2 ; * * , N - 1. (33) dows, the a. coefficient is the coherent gain of the window. coefficients correspond to the minimum 3-term Usingagradientsearch technique [ 9 ] , wehave found the The listed windows which for 3- and 4-nonzero terms achieve a minimum window which is presented in Fig. 24, another 3-term window
(:)
:( )
:( )
65
3-TellIl (-67 d ) B
3-Term (-61 dB)
4 -Term (-92 dB)
.40217 a0 0.44959 0.35875 0.42323 0.49703 a1 0.49364 0.48829 0.49755 a2 0.07922 0.14128 0.09392 0.05677 ___ ---0.00183 0.01168 a3
4-TUIQ windows (parameterized on a) which were the starting condi(-74 dB) tions for the gradient minimbation which leads t o the Black-H ri - ars windows. The opthization.-starhg with these coefficients hasvirtually no effect on the main4obe characteristics but does drive down the sidelobes approximately 5 dB.
F. ConstructedWindows Numerous investigators have constructed windows as prod(to establish another data point in Fig. 12), the minimum 4 - ucts, as sums, as sections, or as convolutions of simple functerm window (to also establish a data point in F'ii. 12), and tions and of other simple windows. These windows have been another 4-term window which is presented in Fig. 25. The constructedfor certain desirable features,not the least of particular 4-term window shown is one which performs well which is the attraction of simple functions for generating the in a detection example described in Section VI (see Fig. 69). window terms. In general, the constructed windows tend not The parameters of these windows are listed in Table I. Note in to be good windows, and occasionally are very bad windows. particular where the Blackman and the Blackman-Harris win- We have already examined some simple window constructions. dows reside in Fig. 12. They are surprisingly good windows The Fejer (Bartlett) window, for instance, is the convolution for the small number of terms in their trigonometric series. of two rectangle windows; the Hamming windowis the sum of Note, if we were t o extend the line connecting Blackman- a rectangle and a Hanning window; and the cos4(X) window the Harris family it would intersect the Hamming window which, is the product of two Hanning windows. We will now examine in Section V-D ,we noted is nearly the minimum sidelobe level other constructed windows that have appeared in the literal l i ture. We wpresent them so they are available for compari2-term Blackman-Harris window. a Blackman- son. Later we will examine windows constructed in accord We also mention that good approximation to the Harris 3- and 4-term windows can be obtained as scaled with some criteria of optimality,(see Sections VG, H, I, and s samples of the Kaiser-Bessel window's transform ( e Section J). Each window i identified only forthe f i t e Fourier transse form. A simple shift of N/2 points and right end-point deleV-H). We have used this approximation t o construct 4-term windows for adjustable bandwidth convolutional filters as tion will supply the DIT version. The significant figures of performance for these windows are also found in Table I. reported in [ 101. This approximation is defined as I ) Riesz (Bochner, Panen) Window [ I 1j : The Riesz window, identified as
uo = C
bo
a, = 2 ,
bm
C
m = 1 , 2 , (3).
(34)
a. = 0.40243, u1 = 0.49804, a2 = 0.09831,
when Q = 3.0 are and a3 = 0.00122. Notice how close theseterms are tothe selected 4-term Blackman-Harris (-74 dB) window. The window defined by thesecoefficients is shown in Fig. 26. Like theprototype from which it came (the Kaiser-Bessel with CY = 3.0), this window exhibits sidelobes just shy of -70 dB from the main lobe. On the scale shown, the two are indistinguishable. The parameters of this window are also listed in Table I and the window is entered in Fig. 12 as the "4-sampleKaiserBessel." It was these 3- and 4-sample Kaiser-Bessel prototype The 4 coefficients for this approximation
is the simplest continuous polynomial window. It exhibits a discontinuous fr t derivative at boundaries; is the hence its transform falls off like l / d . The window is shown in Fig. 27. The fr t sidelobe is -22 dB fromthe main lobe. This is window is similar t o the cosine lobe (26) as can be demonstrated by examinhgits Taylor series expansion. 2 ) Riemunn Window (121: The Riemann window, defined b Y
is the central lobe of the SINC kernel. This window is con-
66
PROCEEDINGS OF THE IEEE, VOL. 66, NO. 1. JANUARY 1978
-15
-20
-15
-10
1
'i !
i
-5
in
1.15
0 d8
Fa.27.
(a) Riesz window. (b) Log-magnitude of transform.
125
-25
4
-20
Fa.28.
(a) Riemann window. (b) Log-magnitude of transform.
I
I
125
Fig. 29. (a) The de la Vall6-Pouspin window. (b) Log-magnitude of transform.
tinuous, with a discontinuous first derivative at tile boundary. It is similar to theKesz and cosine lobe windows. The Riemann window is shownia Fig. 2.8. 3) de l Vall&Poussin (Jackson, Parzen) Window I I : The a ( ] de la VallB-Poussin window is apiecewisecubiccurve obtained by self-convolving two triangles of half extent or four rectangles of one-fourth extent. It is defined as 11.0-
6[d2
N
[1.0-%],
O<Inl<- N 4
2 [LO - $$,
N -<InI<-.
4
2
(37)
The window is continuous up to its third derivative so that its sidelobes fall off like l/w4. The window is shown in Fig. 29. Notice the trade of.ofmain-lobe wid&-fer-sidelobelevel. Compare this with the rectangle and the triangle. It is a nonnegative window by virtue of its self-convolution construction. 4 ) TukeyWindow [13/: The Tukey window, often called the cosine-tapered window, is best imagined as a cosine lobe of width ( a / 2 ) N convolved with arectanglewindow of width . (1 .O - a / 2 ) N Of course the resultant transform is the product of the two corresponding transforms. The window represents an attempt to smoothly set the data to zero at the boundaries while not significantly reducing the processing gain of the windowed transform. The window evolves from the rectangle to the Hanning window as the parameter a varies from zero t o unity. The family of windows exhibits a confusing array of
67
-25
& I
15
i: t
5
io
is
-25
B
-20
sidelobe levels arising from the productof the two component transforms. The window is defined by
tained by the convolution of two half-duration cosine lobes (26a), thus its transform is the square of the cosine lobe's t a s o m ( e Fig. 16). In the time domain the window can r n f r se be described as a product of a triangle window with a single cycle of a cosine with the same period and, then, a corrective term added to set the f rt derivative to zero at the boundary. rs Thus the second derivative is continuous, and the discontinuity resides in the third derivative. The transform falls off like l/w4. The window is defined in the following and is shown i n Fig. 33:
(38)
The window i shown in Figs. 30-32 for values of Q equal to s 0.25,0.50, and 0.75,respectively. 5 ) Bohman Window [14]: The Bohman window is ob-
ff OQlnlQ-. 2
(39)
68
(a)
@)
Fig. 33. (a) Bohman window. (b) Log-magnitude of transform.
125
(a)
@)
Fig. 34. (a) Poisson window. (b) Log-magnitude of transform (a = 2.0).
T T
125
(a)
(b)
Fig. 35. (a) Poisson window. (b) Log-magnitude oftransform (a = 3.0).
6 ) Poisson Window [12]: The Poissonwindow sided exponential d e f i e d by
is a two-
o g l n I <N -.
2
observed in Table I as a large equivalent noise bandwidth and processing as worst case a large loss. 7 Hanning-Poisson Window: The Hanning-Poisson ) win(40) dow is constructed as the product of the Hanning and the Poisson windows. The family is d e f i e d by
This is actually a familyofwindows parameterized on the variable a. Since it exhibits a discontinuity at the boundaries, w ( n ) = 0.5 the transform can fall off no faster than l/w. The window is shown in Figs. 34-36 for values a equal to 2.0,3.0, (41) 4.0, of and respectively. Notice as the discontinuity atthe boundaries This windowsimilar is tothe Poisson window. The rate of becomes smaller, the sidelobe structure merges intothe asymptote. Also note the verywidemain lobe; this will besidelobefalloff is determined by the discontinuity in the fist
69
1.25
(4
@)
Fig. 36. (a) Poismewindow. (b) Logmagnitude of transform (u = 4.0).
(4
(b)
Fig. 37. (a) Hllming-Pobon window. (b) Log-magnitude of transform (u = 0.5).
(8)
(b)
F g 38. (a) Hanning-Poisson window. (b) Log-magnitude of transform (a = 1.0). i.
derivative atthe on& and is I/&. Noticeas a increases, forcing more of the exponential into the Hanningwindow, the zeros of the sidelobe structure disappear and the lobes merge intotheasymptote. This window isshown in Figs. 37-39 for values of a equal to 0.5, 1.O, and 2.0, respectively. Again note thevery large main-lobe width. 8) Cauchy (Abel, Poisson) Window ( 1 S J : The Cauchy window is a family parameterized on a and defined by
w(n)=
Cauchy window is a two-sided exponential (see Poisson windows), which when presented on a log-magnitude scale is essentially an isoscelestriangle. This causes the window t o exhibit a very wide mainlobe and to have a large ENBW.
G. Gaussian or Weiersfrass Window ( I S ] Windows are smooth positive functions with tall thin (i.e., concentrated) Fourier transforms. From the generalized uncertainty principle, we know we cannot simultaneously concentrateboth a signal and its Fourier transform. If our measure of concentration is the mean-square time duration T and the mean-square bandwidth W , we know all functions satisfy the inequality of
1.0+
I-$.[
1
2,
N O<lnl<-. 2
(42)
The window is shown in Figs. 40-42 for values of a equal to 3.0,4.0, and 5.0, respectively. Note the transform of the
Tw>-
4n
(431
70
PROCEEDINGS OF THE IEEE, VOL. 66, NO. 1, JANUARY 1978
T 1
1m .
'fiI'"
(b)
O-
(4
F .4 . (a) i 0
Chuchy window. ( ) Log-magnitude of transform (a = 3.0). b
(8)
9)
(a) Cauchy window. (b) Log-magnitude of tr8nafom (a = 4.0).
Fii. 41.
with equality being achieved only for the Gaussian pulse [ 161. a Thus the Gaussian pulse, characterized by minimum timebandwidth product, is a reasonable candidate for a window. When we use the Gaussian pulse as a window we have to truna t e or discard the tails. By restricting the pulse to be f i t e length, the window no longer is minimum time-bandwidth. If the truncation point is beyond the threesigma point, the error should be small, and the window should be a good approximation to minimum time-bandwidth. T h e Gaussian window is defmed by
Dirichlet kernel as indicated in
Q
(44b)
The transform is the convolution of a Gaussian transform with
This window is parameterized on a, the reciprocal of the standard deviation, a measure of the width of itsFourier transform. Increased a will decreasewith the width of the window and reduce the severity of the discontinuity at the boundaries. This wiU result in an increased width transform
71
1.25
1
I
l.m
-1
8
7
(4
(b)
Fig. 42. (a) Cauchy window. (b) Log-magnitude of transform (a = 5.0).
0d l
T '
1.25
-m
(a)
(b)
Fig. 43. (a) Gaussian window. (b) Log-magnitude of transform (a = 2.5).
T 1
1.25
(a)
(b)
Fig. 44. (a) Gaussianwiudow. (b) Log-magnitude of transform (a = 3.0).
main lobe and decreasedsidelobelevels. The window _is presented in Figs. 43, 44, and 45 for values of a equal to 2.5, 3.0, and 3.5, respectively. Note the rapid drop-off rate of sidelobe level in the exchange of sidelobe level for main-lobe width. The figures of merit forthis window are listed in Table I.
~
H. Dolph-Chebyshev Window [ I 7J Following the reasoning of the previoussection, we seek a windowwhich, for a known finiteduration,in some sense exhibits a narrow bandwidth. We now take alead from the antenna design people who havefaced and solvedasimilar problem. The problem is t o illuminate an antenna of finite
aperture to achievea n m a w main-lobe-beam pattern while simultaneously restricting sidelobe response. (The antenna designercallshisweighting procedure shading.) The closedform solution to the minimummain-lobe width for agiven sidelobe level is the Dolph-Chebyshev window (shading). The continuous solution to the problem exhibits impulses at the boundaries which restricts continuous realizations to approximations (the Taylor approximation). The discrete or sampled window is not so restricted, and the sohtion can he implemented exactly. The relation T,(X)= cos (ne) describes a mapping between the nth-order Chebyshev (algebraic) polynomial and the nthorder trigonometric polynomial. The Dolph-Chebyshev
72
I
I
I
115
1.m
F .45. e
(a) Gaussisnwindow. (b) Log-rmgnitude of M m(a = 3.5).
m
125
0-
i5
-.
@)
(4
Fw.46.
(a) Dolph-Chebyslm window. (b) Lmg-magrdude of transform (a = 2.5).
T
I
-20
-40
-*
(a)
.
0
@)
F .47. i
(a) Dolph-Chebylev window. (b) Log-magnitude of transform (a = 30. .)
window is defined with this mapping in the followingequation, in terms of uniformly spaced samples of the window's Fourier transform,
cos
and cos-1 (X) =
[Ncos-1 [o cos
cosh [Ncosh-I
W(k)= (- 1)k
(*;)]I
@)I
'
1"
- tan-' [ X / ~ E K F I ,
IX I G 1.O
I
of where
'I V
(4s)
To obtain the corresponding window time samples w(n), we simply perform a DFT on the samples W ( k ) and then scale for unity peak amplitude. The parameter Q represents the log of the ratio of main-lobe level t o sidelobe level. Thus a value a equal to 3 0 represents sidelobes 3 0 decades . from .down the main lobe, or sidelobes 60.0 dB below the main lobe. The (-l)& alternates the sign ofsuccessive transform samples to reflect the shifted origin in the time domain. The window is
73
(a)
(b)
Fig. 48. (a) Dolph-Chebyshev window. (b) Log-mn.gnitude of transform (u = 35. .)
Fg 49. i.
(a) Dolph-Chebyshev window. (b) Log-magnitade of M
o m (u = 4.0).
presented in Figs. 46-49 for valuesof a equal to 2.5, 3.0, The parameter nu is half of the timebandwidth product. The 3.5, and 4.0, respectively. Note the uniformity of the sidelobe transform is approximately that of hs structure; almost sinusoidal! It is t i uniform oscillation N (46b) which is responsible forthe impulses in the window. w(e)G - IdSP - (N6/il21. zo(m) Ja2n2 - ( ~ e / 2 ) 2 I. Kaiser-Bessel Window [ l 8 ] This window is presented in Figs. 50-53 for values of Q equal Let us examine for a moment the optimality criteria of the to 2.0, 2.5, 3.0, and 3.5, respectively. Note thetradeoff last two sections. In Section V-G we soughtthefunction between sidelobe level and main-lobe width. with minimum time-bandwidth product. We know this to be the Gaussian. In Section V-H we sought thefunctionwith J. hrcilon-Temes Window [21] restricted time duration, which minimized the main-lobe We now examine the last criterion of optimality for a winwidth for a given sidelobe level. We now consider a similar dow. We have already described the Slepian, Pollak, and problem. For a restricted energy, determine the function of Landau criterion. Subject to the constraints of fmed energy restricted time duration T which maximizes the energy in the and fved duration, determine the function which maximizes bandof frequencies W. Slepian, Pollak,andLandau [ 191 , the energy in the band of frequencies W. A related criterion, [20] have determined this function as a family parameterized subject to the constraints of fiied area and fmed duration, is over the time-bandwidth product, the prolate-spheroidal wave to determine the function whichminimizes the energy (or functions of order zero. Kaiser has discovered a simpleapthe weighted energy) outside the band of frequencies W. This proximation to these functions in terms of the zero-order is a reasonable criterion since we recognize that the transform modified Bessel function of the fiist kind. The Kaiser-Bessel of a good window should minimize the energy it gathers from window is defined by frequencies removed from its center frequency. Till now, we have been responding to t i goal by maximizing the concenhs tration of the transform at its main lobe. A closed-form solution of the unweighted minimum-energy criterion has not been found. A solution defined as an expansion of prolate-spheroidal wave functions does exist and it is of the form shown in
14
135
1.00
-1 -20
-15
-10
-5
Fg 50. i.
(a) KPirer-Jhsd window. (b) Logmagnitude of transform (u = 2.0).
.+.
1m .
(a)
@)
(a) Kaiser-Bessel window. (b) Logmagnitude of transform (a = 2.5).
Fg 51. i.
(a)
(b 1
Fig. 52. (a) Kaiser-Bessel window. (b) Log-magnitude of transform (a = 3.0).
1.25 1.00
I
!!
L !
Fig. 53. (a) Kaiser-Bessel window. (b) Log-magnitude of transform (u = 35. .)
75
1
0
I *
125
1.m
F .54. a
(a) B.r&n-Tma
window. ( ) Log-magnhde of transform (u = 3.0). b
I I
1.25
lm
(a)
(b)
Fa.55.
(a) Barcilon-Temes window. (b) Log-magnitude of tr8asform (u = 3.5).
Ti. 56. (a) Barcilon-Temes window. ( ) Lag-nugnhdeof transform (u = 4.0). b
Here the A,,, is the eigenvalue corresponding to the associated Like the Dolph4%ebyshev window, the Fourier transform is prolate-spheroidal wave function I$42Jx, y ) I, and the tra is more easily defined, and the window timesamples are obthe selected half time-bandwidth product. The summation tained by an inverse DFT and an appropriate scale factor. The converges quite rapidly, and is often approximated by the first transform samples are defined by term or by the first two terms. The first term happens t o be thesolution of the Slepian, Pollak, and Landau problem, which we have already examined as the Kaiser-Bessel window. A cos y(k)l + B sin ~ ( k ) , ] A closed-form solution of a weighted minimum-energy W(k)= (- 1)k criterion, presented in the following equation has been found [C+ABl + 1-01 by Barcilon and Temes:
r$ [p?]'
(49)
This criterion is one which is a compromise between the DolphChebyshev and the Kaiser-Bessel window criteria.
16
C = cosh-' ( 1OQ) @= cosh [ i c ]
y(k)=Ncos-'
[ B cos
(41
-Q
1
c
T
(See also (45).) This window is presented in Figs. 54-56 for values of equal to 3.0, 3.5, and 4.0, respectively. The mainlobe structure is practically indistinguishable from the KaiserBessel main-lobe. The f i e s of merit listed on Table Isuggest that for the samesidelobelevel, this window doesindeed reside between the Kaiser-Bessel andthe Dolph-Chebyshev windows. It is interesting to examine Fig. 12 and note where t i window is located hs with respect tothe Kaiier-Bessel window; striking similarity in performance!
-ea
VI. HARMONIC
ANALYSIS
We now describe a simple experiment which dramatically demonstrates the influence a window exerts on the detection of a weak spectral line in the presence of a strong nearbyline. If two spectral l n s reside in DFT bins, the rectangle window ie allows each to be identified with no interaction. ,To demonstrate this, consider the signal composed of two frequencies 10 f , / N and 16 f,/N (corresponding t o thetenth and the sixteenth DFT bins) and of amplitudes 1.0 and 0.01 (40.0 dB separation), respectively. The power spectrum of this signal obtained by a DFT is shown in Fig. 57 as a linear interpolation between the DFT output points. We now modify the signal slightly so that the larger signal resides midway between two DFT bins; in particular, at 10.5 f s / N . The smaller signal still resides in the sixteenth bin. The power spectrum of this signal is shown in Fig. 58. We note that the sidelobe structure of the larger signal has completely swamped the main lobe of the smaller signal. In fact, we know (see Fig. 13) that the sidelobe amplitude of the rectangle window at 5.5 bins from the center is only 25 dB down from the peak. Thus the secondsignal ( 5 . 5 bins away) could not be detected because it was more than 26 dB down, and hence, hidden by the sidelobe. (The 26 dB comes from the 25dB sidelobe level minus the 3.9dB processing loss of the window plus3.0dB for a high confidence detection.) We also note the obvious asymmetry around the main lobe centered at 10.5 bins. This is due to the coherentaddition of the sidelobe structures of the pair of kernels located at the plus and minus 10.5 bin positions. We are observing the self-leakage between the positive and the negative frequencies. Fig. 59 is the power spectrum of the signal pair, modified so that the large-amplitude signalresides at the 10.25-bin position. Note the change in asymmetry of the main-lobe and the reduction in the sidelobe level. We still can not observe the second signal located at bin position 16.0. We now apply different windows to the two-tone signal to demonstrate the difference in second-tone detectability. For some of the windows, the poorer resolution occurs when the largesignal is at 10.0 bins rather than at 10.5 bins. We will always present the window with the large signal at the location corresponding to worst-case resolution. The first window we apply is the triangle window (see Fig. 60). The sidelobeshavefallenby a factor of two over the rectangle windows' lobes (e.g., the -35dB level has fallen to -70 dB). The sidelobes of the larger signal have fallen to approximately -43 dB at the second signal so that it is barely
-ea
T +
1
I
I k
Fii. 56.
Rectangle window.
-60
*
I
Fig. 5 9 . Rectangle window.
O?
-m
1 i
I\
Fii. 60.
Triangle window.
il
detectable. If there wereany noise in the signal, the second tone would probably not have been detected. Thenext windows we apply are the cosa(x) family. For the cosine lobe, a = 1.O, shown in Fig. 6 1 we observe a phase cancellation in the sidelobe of the large signal located at the smallsignal position. This cannot be considered a detection. We alsosee the spectral leakage of the main lobe over the frequency axis. Signals below this leakage level would not be detected. With a = 2.0 we have the Hanning window, which is
77
Amd. 1.m
not
I n
Fig. 61. Cos (nn/N) window.
Fig. 64. Cos (nn/N) window.
-m
jot l i
*
/I
Fig. 66. Blockman window.
-40
-60
Fig. 63. Cos ( n f f / N window. )
presented in Fig. 62. We detect the second signal and observe a 3.0-dB null between the two lobes. This is still a marginal detection. For the cos3(x) window presented in Fig. 63, we detect the second signal and observe a 9.0dB null between the lobes. We also see the improved sideloberesponse. Finally for the cos4(x) window presented in Fig. 64, we detect the secondsignalandobserve a 7.0dB null between the lobes. Here we witness the reducedreturn for the trade between sidelobe level and main-lobe width. In obtaining further reduction in sidelobe level we have caused the increased mainlobe widthto encroach uponthe second si nal g . We next apply the Hamming window and present the result in Fig. 65. Here we observe the second signal some 35 dB down,approximately 3.0dBover the sidelobe responseof the large signal. Here, too, we observe the phase cancellation and the leakage between the positive and the negative frequencycomponents. Signals morethan 50 dB down would not be detected in the presence of the larger signal. The Blackman window is applied next andwe see the results in Fig. 66. The presenceof the smaller amplitude kernel is nowvery apparent. There is a 17-dB null between the two signals. The artifact at the base of the large-signal kernel is
Fig. 67. Exact Blackman window.
the sidelobe structure of that kernel. Note the rapid rate of falloff of the sidelobe leakage has confined the artifacts to a small portion of the spectral line. We next apply the exact Blackman coefficients and witness the results in Fig. 67. Again the second signal is well defiied with a 24dB null between the two kernels. The sidelobe structure of the larger kernel now extends over the entire spectral range. This leakage is not terribly severe as it is nearly
F .71. s
Riesz window.
=-I 1
j
F .72. Riemmn window. a
-B
F .73. a
6OdB down relative to the peak. There is another. small artifact at 5OdB down on the low frequency side of the large kernel. This is definitely a single sidelobe of the large kernel. This artifact is essentially removed by the minimum 3-term Blackman-Harris window which we see in F i i 68. The null between the two signal main lobes is slightly smaller, at approximately 20 dB. Next the 4-term Blackman-Harris window is applied to the signal and we see the results in Fig. 69. The sidelobe structures are more than 7OdB down and as such are not obserrred on t i scale. The two signal lobes are well defined with hs approximately a 19dB null between them. Now we apply the 4-sample Kaiser-Bessel window to the signal and see the results in Fig. 70. We have essentially the same performance as with the 4-term Blaclanan-Harris window. The only observable difference on this scale is the small sidelobeartifact 68 dB down on the low frequency side of the lnrge kernel. This group of Blackmanderived windows perform admiraMy wen for their s m l c t . ipiiy The Riesz window is the first of our constructed windows e and is presented in Fig. 7 1. W have not detected thesecond signal but we do observe iis affect as a 20.ndl due
de t V&Poussin
window.
to the phase cancellation of a sidelobe in the large signal's kernel. The result of a Riemannwindow is presented in Fig. 72. Here, too, we have no detection of the second signal. We do have a small null due to phase cancellation at the,second signal. We also have a large sidelobe response. Thenext window, thede la VaIli-Poussin orthe selfconvolved triangle, is shown in Fig. 73. The second signal is easily found and the power spectrum exhibits a 16.0-dB null. An artifact of the window (its lower sidelobe) shows up, however, at the Blh DFT bin as a signal approximately 53.0 dB down. See Fig. 29. The result of applying the Tukey family of windows is presented in F i . 74-76. In Fig. 74 (the 25-percent taper) w see the lack of second-signal detection dueto thehigh sidee lobe structure of the dominant rectangle window. In Fig. 75 (the 50-percent taper) we obseme a lack of seconddgnal detection, with the second signal actually filling in one of the nulls of the hfft sgas kernel. In Fig. 76 (the 76-percent inl' taper) we witness a marginal detection in the still high sidelobes of the larger s g a . This is still an unsatisfying window inl because of the artifacts.
79
I
O I h
\i
i
\A
A & l ~ W i i 9 b ~ ~
Fig. 74. Tukey (25-percent cosine taper) window.
Fig. 78. Poisson window (a = 2.0).
F g 75. Tukey (50-percent cosine taper) window. i.
Fig. 79. Poisson window (a = 3.0).
-20
-20
-a
-40
-e3
-M
O
Fig. 76. Tukey (75-percent cosine taper) window.
&
Fa.80.
Poisson window (a = 4.0).
The Bohman construction window is applied and presented in Fig. 77. The second signal has been detected and the null between the two lobes is approximately 6.0dB. This is not bad, but we can still do better. Note where the Bohman window resides in Fig. 12. The result of applying the Poisson-window family is presecond signal is not detected for sented in Figs.78-80.The any of the selected parameter values due t o the highsidelobe
levelsof the largersignal. We anticipatedthispoor performance in Table I by the large difference between the 3.0 dB and the ENBW. The result of applying the Hanning-Poisson family of windows is presented in Figs. 81-83. Here, too, the second signal is either not detected in the presenceof thehighsidelobe structure or thedetection is bewildered by the artifacts. The Cauchy-family windows been have applied and the results are presented in Figs. 84-86. Here too we have a lack of satisfactory detection of the second signaland the poor sidelobe response. This was predicted by the large difference between the 3.0dB and the equivalent noise bandwidths as listed in Table I. We now apply the Gaussian family of windows and present the results in Figs. 87-89. The second signal is detected in all three figures. We note as we further depress the sidelobe structure t o enhance second-signal detection, the null deepens to approximately 16.0 dB andthen becomes poorer as the main-lobe width increasesand starts t o overlap the lobe of the smaller signal. The Dolph-Chebyshev family ofwindows is presented in Figs. 90-94. We observe strong detection of the second signal
PROCEEDINGS OF THE IEEE. VOL. 66, NO. 1. JANUARY 1978
-m
b l & 2 b 2 & Q X X ? b h i G
F 85. =
Caucby window (u = 4.0).
bl&ab;P;D;Dd7AbSirE
Fii. 86.
Csuchy window (a = 5.0).
FFTM
105
md.
1.00 QO1
SllrZ
1.
lU.0
-zD
n
Fii. 88. Gaussian window (a = 3.0).
-a
-60
ng.84. Cauchy window (a = 3.0).
-a
in all cases, but it is distressing t o see the uniformly high sidelobe structure. Here, we again see the coherent addition of the sidelobes from the positive and negative frequency kernels. Notice that the smaller signal is not 4OdB down now. What we are seeing is the scalloping loss of the large sg as maini nl' lobe being sampled off of the peak and beiug refereaced as zero dB. Figs. 90 and 91 demonstrate the sensitivity of the sidelobe coherent addition to main-lobe position. In Fig. 90 the larger signal is at bin 10.5; in Fig. 91 it is at bin 10.0.
Note the difference in phase cancellation near the base of the large s g a . Fig. 93, the 7MB-sidelobe window, exhibits an inl 18-dB null between the two main lobes but the sidelobes have added constructivdy (along with the scalloping loss) to the -62.O-dB level. In Fig. 94, we see the 80-dB sidelobe window exhibited sidelobes below the 70-dB level and still managed to hold the null between the two lobes to approximatley 18.0 dB. The Kaiser-Bessel family is presented in Figs. 95-98. Here,
ANALYSIS
81
Fig. 89. Gaussian window (a = 3.5).
Fig. 93. Dolph-Chebyshev window (a = 3.5).

FFT Sin
Amw I m 0.01
s,pn*
I S # p n * 2.
10.5
16.0
Fig. 9 1 . Dolphzhebyshevwindow (a = 2.5).
Fig. 95. Kaiser-Bessel window (a = 2.0).

FFTBm
Siqd 1. Siqd 2 10.5 16.0
Am#, 1m
00 .1
-eo i
too, we have strong second-signal detection. Again, we see the effect of trading increased main-lobe width for decreased sidelobe level. The null between the two lobes reaches a maximum of22.0dBas the sidelobe structure falls and then becomes poorer with further sidelobe level improvement. Note that this window can maintain a 20.0-dB null between the two signal lobes and still hold the leakage to more than 70 dB down over the entire spectrum. Figs. 99-101 present the performance of the BarcilonTemes window. Note the strong detection of the second signal.
There areslightsidelobe artifacts. The window can maintain a 20.0dB null between the two signal lobes. The performance of this window is slightlyshy of that of the Kaiser-Bessel window, but the two are remarkably similar.
VII. CONCLUSIONS
We have examined some classic windows and some windows whichsatisfysome criteria o optimality.Inparticular, f we Jmve dacribeil their effects on the problem of generalhar-
82
FFIBin

W.
0.01
sm +
1.
105
1.m
jilru 2.
16.0
-m
-m
-40
-a
-8u
-60
,
o l b ~ i a D & & S b o & a m

Fig. 97. Kaiser-We1 window (a = 3.0).
+ k
,
~
,
O
,
~
,
~
,
~ ~
,
~
, t
~ O L I O
Fig. 100. Bardon-Temes window (a = 3.5).
I t
,
Ib
1
i n
io
lo
k&J
70
l %
Fig. 101. Barcilon-Temes window (a = 4.0).
' ;
I , I l ' !
, k
O 9 0 l W
Fig. 99. Barcilon-Temes window (a = 3.0).
monic analysis of tones in broadband noise and of tones in the presence of other tones. We have observed that when the DFT is usedasa harmonic energy detector, the worstcase processing loss due tothe windows appears to be lower bounded by3.0dB and (for good windows) upper bounded near 3.75 This dB. suggests that the choice of particular windows has very little effect on worst case performance in DFT energy detection. We have concluded that a good performance indicator for the window is the difference between the equivalentnoise bandwidth and the 3.0dB bandwidth normalizedby the 3.0-dB bandwidth. The windowswhichperform well (as indicated in Fig. 12) exhibit values forthis ratio between4.0 and 5.5percent. The range of thisratio for the windows listed in Table I is 3.2 to 22.9 percent. For multiple-tone detection via the DFT, the window employed does have a considerable effect. Maximum dynamic range of multitone detection requires the transform of the window to exhibit ahighly concentrated centrallobe with very-low sidelobe structure. We have demonstrated that many classic windows satisfy this criterion with varying
degrees of success and some not at all. We have demonstrated theoptimal windows(Kaiser-Bessel,Dolph-Chebyshev, and Barcilon-Temes) and the Blackman-Hamswindows perform best indetection of nearby tones of significantly different amplitudes. Also for the same dynamic range, the three optimalwindows and the Blackman-Harriswindow are roughly equivalent with the Kaiser-Bessel and the Blackman-Harris, demonstrating minor performance advantages over the others. We note that while the Dolph-Chebyshev window appears to be the best window by virtue of its relative position in Fig. 12, the coherent addition of its constant-levelsidelobes detracts from its performance in multi tone detection. Also the sidelobe structure of the Dolph-Chebyshev window exhibits extreme sensitivity to coefficient errors. This would affect its performance in machines operating with fixed-point arithmetic. This suggests that the Kaiser-Bessel or the BlackmanHarris window be declared the top performer. My preference is the Kaiser-Bessel window. Among other reasons, the coefficients are easy to generate and the trade-off of sidelobe level as a function of time-bandwidth product is fairly simple. For many applications, the author would recommend the 4sample Blackman-Hams (or the 4-sample Kaiser-Bessel) window. These have the distinction of being defined by a few easily generated coefficients and of being able to be applied as a spectral convolution after the DFT. We have called attention to a persistent 'error in the application of windows when performing convolution inthe frequency domain, i.e., the omission of the alternating signs on the window sample spectrum to account for the shifted time origin. We have also identified and clarified a source of confusion concerning the evenness of windows under the DFT. Finally, we comment that allof the conclusions presented about window performance inspectral analysis are also applicable to shading for arrayprocessing of spatial sampled data, including FFT beamfonning.
HARRIS: USE O F WINDOWS FOR HARMONIC ANALYSIS
83 [ 4 ] C. Lanczos, Dircourse on Fourier Series. New York: Hafner Publishing CO., 1966, ch. 1, pp. 29-30. [ 51P.D. Welch, The use of fast Fourier transform for the estimation of power spectra: A method based on time averaging ovir short, modified periodograms, ZEEE Ta s Audio r n. Electroacourt.,vol. AU-15, pp. 70-73, June 1967. [ 6 ] J. R. Rice, The Approximation of Functions, Vol. I. Reading, MA: Addison-Wesley, 1964, ch. 5.3, pp. 124-131. [ 7 ] R. B. Blackman and J. W. Tukey, The Measurement of Power Spectra. New York: Dover, 1958, appendix B.5, pp. 95-100. [ 8 ] L. Fejer, Untersuchunger uber Fouriersche Reihen,Mat. Ann., 58, PP. 501-569,1904. 191 L. R. Rabiner, B. Gold, and C. A. McGonegal, An approach t o the approximation problem nonrecursive for digital fdters, IEEE Trans. Audio Electroacoust., vol. AU-18, pp. 83-106, June 1970. [ l o ] F. J. Harris, High-resolution spectral analysis with arbitrary spectral centers and adjustable spectral resolutions, J. Comput. Elec. Eng.,vol. 3, pp. 171-191,1976. [ 11] E. Parzen,Mathematicalconsiderationsintheestimationof spectra, Technometrics,vol. 3, no. 2, pp. 167-190, May 1961. [ 121 N. K. Bary, A Treatire onTrigonometricSeries, Vol. I. New York: Macmillan, 1964, ch. 1.53, pp. 149-150,ch. 1.68, pp. 189192. [ 131 J. W. Tukey, An introduction to the calculations of numerical spectrum analysis, in Spectral Analysis of Time Series, B. Harris, Ed. New York: Wiley, 1967, pp. 25-46. [14) H. Bohman, Approximate Fourier analysis of distribution vol. 4, 1960,pp. 99-1 57. functions, Arkiv Foer Matematik, [ l S ] N. I. Akhiezer, Theory of Approximation. New York: Ungar, 1956, ch. IV.64, pp. 118-120. Prentice(161 L. E. Franks, Signal Theory. Englewood Cliffs, NJ: Hall, 1969, ch. 6.1,pp. 136 -1 37. [ 171 H. D. Helms, Digital filterswithequirippleorminimaxresponses, ZEEE Trans. AudioElectroacourt., vol. AU-19,pp. 87-94, Mar. 1971. [ 181 F. F. Kuo and J. F. Kaiser, System Analysis by Digital Computer. New York: Wiley, 1966, ch. 7, pp. 232-238. [ 191 D. Slepianand H. Pollak,Prolatespheroidal wave functions, BeU Tel. Syst. J . , vol. 40, Fourier analysis and uncertainty-I, pp. 43-64, Jan. 1961. [20] H. Landauand H. Pollak,Prolatespheroidal wave functions, Fourier analysis and uncertainty-11, EeU Tel. Syst. J . , vol. 40, pp. 65-84, Jan. 1961. I211 V. Barcilon and G. Temes, Optimum impulse response and the Van Der Maas function,ZEEE Trans. Circuit Theory, vol. CT-19, pp. 336-342, July 1972.
APPENDIX THE EQUIVALENCE WINDOWING IN THE TIME OF DOMAINTO CONVOLUTIONI N THE FREQUENCY DOMAIN
Let
+f ( t )=/
-0
F ( a ) exp ( - j u t ) d a / 2 n
and
+Nf 2
W ( a )=
n=-Nf2
w ( n T ) exp (+janT).
Then
+-
~ , ( a= )
becomes
+-
n=--
w ( n T ) f ( n T )exp ( + j a r ~ ~ )
F,(w) =
nr-w
+w(nT)/
-00
F ( x ) exp ( - j x n T ) dx/2n
exp ( + j a n T )
+F(x)
w ( n T ) exp [ + j ( a- x ) n T ] d x / 2 n
or
F,(a) = F ( a ) * W ( a ) .
REFERENCES
[ 11 C.W. Helstrom, Statistical Theory of Signal Detection, 2nd ed. New York: Pergamon Press, 1968, Ch. IV, 4, pp. 124-130. [ 2 ] J. W. Cooley, P.A. Lewis, and P.D. Welch, The finite Fourier transform, IEEE Trans. Audio Electroacoust., vol. AU-17, pp. 77-85, June 1969. [ 31 J. W. Wozencraft and I. M. Jacobs, Principles of Communication Engineering. New York: Wiley, 1965, ch. 4.3, pp. 223-228.
BIBLIOGRAPHY -ADDITIONALGENERALREFERENCES
R. B. Blackman, Data Smoothing and Prediction. Reading, MA: Addison-Wesley, 1965. D. R. Brillinger, Time Series Data Analysis and Theory. New York: Holt, Rinehart, and Winston, 1975., D. Gingras, Time series windows for improving discrete spectra estimation, Naval Undersea ResearchandDevelopmentCenter,Rep. NUC TN-715, Apr. 1972. F. J. Harris, Digital signal processing, Class notes, San Diego State Univ., 1971. G . M. Jenkins, General considerations in the estimation of spectra, Technometric& vol. 3, no. 2,pp. 133-166, May 1961.

Use of Windows For Harmonic Analysis With DFT

Încărcat de

Informații document

Descriere originală:

Titlu original

Drepturi de autor

Formate disponibile

Partajați acest document

Partajați sau inserați document

Opțiuni de partajare

Vi se pare util acest document?

Este necorespunzător acest conținut?

Drepturi de autor:

Formate disponibile

Use of Windows For Harmonic Analysis With DFT

Încărcat de

Drepturi de autor:

Formate disponibile

PROCEEDINGS OF THE IEEE, VOL. 6 6 , NO.

related to sampled windows for DFT's.

11. HARMONIC ANALYSIS OF FINITE-EXTENT DATA AND THE DFT

U.S. Government work not protectedby U.S. copyright

PROCEEDINGS OF THE IEEE, VOL. 6 6 , NO. 1 , JANUARY 1978

Fig. 1 . N samples of an even function taken over an NT second interval.

Fig. 3. DFT sampling of finite Fourier transform of a DFT even sequence.

HARRIS: USE OF WINDOWSFORHARMONIC

Fig. 4. Periodic extension of sinusoid not periodic in observation interval.

IV. WINDOWS AND FIGURES

f(nn(-juknT), exp Neven

PROCEEDINGS OF THEIEEE, VOL. 66, NO. 1, JANUARY 1978

Fig. 6. Graphical interpretation of equation (6). Window visualized as a spectral Nter.

Fig. 7 . Equivalent noise bandwidth of window.

Peak Power Gain = W 2 ( 0 ) =

HARRIS: USE OF WINDOWS FOR HARMONIC ANALYSIS

TABLE I WINDOWS AND FIGURES MERIT OF

U=0.25 U=0.50 U = 0.75

1.10 1.22 -19 1.26 1.79 1.30

1.01 0.75 1.15 1.31 1.71 1.21 1.45 1.75

2.96 2.24 1.73 1.02 2.09 1.46 1.03

3.39 3.1 1 3.07 3.54

1.38 1.57 1.80 2.38 1 2.08 2.58

-19 -24 -31

1.11 0.87 1.48 1.71 1.36

2.14 2.30 2.65

-31 -35 -30

1.76 2.06 1.39

3.83 2.20 4.28 1.13 2.53

a=2.5 U = 3.0 U = 3.5

-42 -55 -69

1.33 1.55 1.79

1 1.64 1.25 0.94

DOLPHa=2.5 CHEEYSHEV a = 3.0 U = 3.5 U =4.0

1.39 1.51 1.62 1.73 15 0 1.65 1.80

1.70 1.44 1.25 1.10 1.46

1.02 1.93 0.89 1.34 1.18

a = 3.0 u = 3.5 U=4.0

-53 -58 -68 -51

1.56 1.67 1.77 1.57 1.73 1.71

1.52 1.68 1.66 1.90 1.56 1.74 1.74

62.7 56.7 57.2

BLACKMAN M I N I M U M 3-SAMPLE BLACKMAN-HARRIS ' M I N I M U M 4-SAMPLE BLACKMAN-HARRIS

1.10 1.13 0.83 1.27 1.03

0.36 0.45 0.40 0.40

2.00 1.61 1.79 1.BO

*REFERENCE POINTS FOR DATA ON FIGURE 1 2 - NO FIGURES TO MATCH THESE WINDOWS.

PROCEEDINGS OF THE IEEE,

NO. 1 , JANUARY 1978

Windowed Sequences (Overlapped)

Fig. 8. Partition of sequences for nonoverlapped and for overlapped pro-g.

. exp ( - j w k n T ) exp (+jwkrnT)

Fig. 9. Relationship between indices on overlapped intervals.

HARRIS: USE OF WINDOWS FOR HARMONIC ANALYSIS

Fig. 10. Spectral leakage

Fig. 1 1 . Spectral resolution of nearby kernels.

PROCEEDINGS OF THE IEEE,

VOL. 66, NO. 1 , JANUARY 1978

and is shown in Fig. 13. Thesame-windowforaDFT

HARRIS: USE OF WINDOWSFORHARMONIC

Fig. 13. (a) Rectangle window. (b) Log-magnitude of transform.

(a) Fig. 14. (a) Triangle window. (b) Log-magnitude of transform.